[flang-commits] [flang] [flang][OpenMP] Do not emit barrier after SECTIONS with LASTPRIVATE and NOWAIT (PR #216018)

[via flang-commits]

https://github.com/ejose02 created https://github.com/llvm/llvm-project/pull/216018

Fixes #192907

Flang inserted an explicit omp.barrier after SECTIONS when both lastprivate and nowait were present, which serialized all threads at the end of the construct and prevented nowait from taking effect.

Limit the barrier to lastprivate(conditional:) cases, where reduction results must be finalized before the post-sections copy-back. Regular lastprivate + nowait now matches wsloop lowering behavior.

>From 980e0247213e1d2fa66b50463441a0d85a63da60 Mon Sep 17 00:00:00 2001
From: ejose <ejose at amd.com>
Date: Thu, 13 Aug 2026 11:26:05 +0000
Subject: [PATCH] [flang][OpenMP] Do not emit barrier after SECTIONS with
 LASTPRIVATE and NOWAIT

Flang inserted an explicit omp.barrier after SECTIONS when both lastprivate and nowait were present, which serialized all threads at the end of the construct and prevented nowait from taking effect.

Limit the barrier to lastprivate(conditional:) cases, where reduction results must be finalized before the post-sections copy-back. Regular lastprivate + nowait now matches wsloop lowering behavior.
---
 flang/lib/Lower/OpenMP/OpenMP.cpp    | 17 +++--------------
 flang/test/Lower/OpenMP/sections.f90 |  1 -
 2 files changed, 3 insertions(+), 15 deletions(-)

diff --git a/flang/lib/Lower/OpenMP/OpenMP.cpp b/flang/lib/Lower/OpenMP/OpenMP.cpp
index a16889aa365d9..83072f0e7be51 100644
--- a/flang/lib/Lower/OpenMP/OpenMP.cpp
+++ b/flang/lib/Lower/OpenMP/OpenMP.cpp
@@ -3902,9 +3902,6 @@ genSectionsOp(lower::AbstractConverter &converter, lower::SymMap &symTable,
   llvm::SmallDenseSet<const semantics::Symbol *> condLpSymSet(
       condLpSyms.begin(), condLpSyms.end());
 
-  // Track whether any non-conditional lastprivate copy-backs were emitted.
-  bool hasNonCondLastprivate = false;
-
   if (!lastprivates.empty()) {
     mlir::Region &sectionsBody = sectionsOp.getRegion();
     assert(sectionsBody.hasOneBlock());
@@ -3926,7 +3923,6 @@ genSectionsOp(lower::AbstractConverter &converter, lower::SymMap &symTable,
         // Skip conditional LP symbols — handled by the reduction path.
         if (condLpSymSet.count(sym))
           continue;
-        hasNonCondLastprivate = true;
         if (const auto *common =
                 sym->detailsIf<semantics::CommonBlockDetails>()) {
           for (const auto &obj : common->objects())
@@ -3941,16 +3937,9 @@ genSectionsOp(lower::AbstractConverter &converter, lower::SymMap &symTable,
   // Perform DataSharingProcessor's step2 out of SECTIONS
   builder.setInsertionPointAfter(sectionsOp.getOperation());
   dsp.processStep2(sectionsOp, false);
-  // Emit barrier when nowait is present and there are lastprivate copy-backs
-  // (either non-conditional or conditional).  The barrier ensures all threads
-  // have completed their work before lastprivate values are read/copied.
-  //
-  // NOTE: The LLVM OpenMP runtime currently imposes an implicit barrier
-  // inside __kmpc_reduce for tree reductions.  If the runtime were modified
-  // to release losing threads early when nowait is specified, we could use
-  // the return value from the tree reduction (case 1 = winner) to let the
-  // winner thread perform the copy-back without a separate barrier.
-  if (clauseOps.nowait && (hasNonCondLastprivate || !condLpSyms.empty()))
+  // Emit barrier when nowait is present and conditional lastprivate reduction
+  // results must be finalized before copy-back.
+  if (clauseOps.nowait && !condLpSyms.empty())
     mlir::omp::BarrierOp::create(builder, loc);
 
   // Copy-back: copy winning values from the shared reduction struct to the
diff --git a/flang/test/Lower/OpenMP/sections.f90 b/flang/test/Lower/OpenMP/sections.f90
index e3155661f35bc..9767c76cf104d 100644
--- a/flang/test/Lower/OpenMP/sections.f90
+++ b/flang/test/Lower/OpenMP/sections.f90
@@ -209,7 +209,6 @@ subroutine lastprivate()
             x = x + 1
 !CHECK: omp.terminator
 !CHECK: }
-!CHECK: omp.barrier
      !$omp end sections nowait
 
 !CHECK: %[[PRIVATE_X:.*]] = fir.alloca i32 {bindc_name = "x", pinned, uniq_name = "_QFlastprivateEx"}